Fault propagation path identification method based on a causal- enhanced bidirectional graph attention network

As modern industrial processes become increasingly large-scale and integrated, the interactions and coupling among process variables grow substantially more complex, making fault-source localization and fault propagation path identification increasingly challenging. Conventional fault-source graphs are typically constructed from statistical correlations or predefined process topologies and therefore fail to capture directional causal information, which may produce redundant or misleading propagation paths. To overcome these limitations, a causal-enhanced bidirectional graph attention network (CE-Bi-GAT) for fault propagation path identification is proposed. Specifically, a path-aware transfer entropy (PATE) method is developed to construct a directed weighted causal graph by integrating k -nearest neighbor candidate screening, conditional transfer entropy estimation, and k -PHop path-aware aggregation. Based on the resulting causal graph, a bidirectional graph attention network is designed to simultaneously model forward fault propagation and reverse diagnostic inference. Furthermore, a reverse attention mechanism is introduced to decode symptom-to-root propagation paths, while a joint optimization objective enhances both root-cause localization and path interpretability. Experimental results on the Tennessee Eastman process demonstrate that CE-Bi-GAT achieves a path-tracing accuracy of 94.6%, a root-cause localization accuracy of 82.5%, and a path significance index of 0.78, consistently outperforming conventional baseline methods and recent graph-attention-based fault diagnosis approaches. Two synthetic ammonia case studies further demonstrate the cross-process generalizability of the proposed method.